A practical log of building a long-lived Telegram AI agent — caching, providers, routing, memory, latency. No benchmarks. Just what actually happened.
Latency gap (2.2s LLM vs 6-11s full agent) and 66% cache hit rate reveal infrastructure layers—not model quality—drive performance. For IT managers: effective agents need telemetry depth and prefix stability—90% caching benchmarks mislead; real ROI is infrastructure engineering where routing costs cascade five layers deep.
Not a researcher. Not a professional dev. Civil engineering background. Started building an AI bot because I wanted to understand what's actually happening inside these systems — not theoretically, just practically.
Wanted an assistant that could live with a conversation instead of treating every message as an isolated API call.
It started as an experiment. Then the experiment grew.
What it turned into
At some point I had:
Six months ago I got tired of switching between apps to talk to AI. ChatGPT in the browser. Claude in...
Build a personal AI agent in Telegram using Bun, Telegraf, SQLite, and Composio. Zero-infrastructure with conversation memory and…
Build a personal AI agent in Telegram using Bun, Telegraf, SQLite, and Composio. Zero-infrastructure with conversation memory and…
Most "AI agent" demos die at the same place: a tweet, a screenshot, a five-minute video. Then the...
A few months ago I started posting on Dev.to about my side projects. It was fun at first — writing...
An autonomous AI agent that wakes once an hour on the five hard-won lessons from running a three-layer agent hierarchy:…
